June 3, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Anthropic confidentially files IPO registration with SEC
Anthropic has submitted a confidential draft S-1 registration statement to the SEC as part of a proposed initial public offering. No pricing or share counts have been set, and the process remains subject to regulatory review and market conditions.
NVIDIA launches Nemotron Ultra (550B) and Cosmos 3 for physical AI
NVIDIA announced two major models: Nemotron Ultra, a 550-billion-parameter (55 billion active) open-weights model that tops US AI benchmarks, and Cosmos 3, a fully open omnimodel for physical AI that natively handles text, image, video, ambient sound, and actions via a mixture-of-transformers architecture.
OpenAI models and Codex now generally available on AWS Bedrock
OpenAI has made its frontier models and Codex generally available on AWS, letting enterprises access OpenAI capabilities through existing AWS security, governance, procurement, and billing workflows. OpenAI's developer cookbook covers production patterns including structured outputs, tool calling, and prompt caching via the Responses API.
Alphabet raising $80 billion in stock sales to fund AI infrastructure
Alphabet plans to raise $80 billion through stock sales to fund AI compute infrastructure, driven by record customer demand. The raise includes $10 billion from Berkshire Hathaway, $30 billion in an underwritten offering, and $40 billion via an at-the-market program.
NVIDIA introduces RTX Spark Arm SoC for AI-focused Windows PCs
NVIDIA unveiled RTX Spark at Computex 2026, an Arm-based SoC for Windows PCs featuring up to 20 CPU cores and a Blackwell GPU with 6,144 CUDA cores, targeting premium laptops and small form factor desktops. Built in partnership with Microsoft, the chip is also designed to run personal AI agents locally.
Perplexity introduces Search as Code for AI-driven search pipelines
Perplexity launched Search as Code (SaC), a new architecture that gives AI models direct control over the search stack via an SDK, letting them assemble custom search pipelines rather than consuming pre-packaged results. The approach improves performance and flexibility over traditional monolithic search systems.
Anthropic API no longer charges for refusals, adds advisor token cap
Anthropic updated Claude API billing so that requests returning stop_reason: "refusal" with no generated output are no longer charged. The advisor tool also now supports a max_tokens parameter to cap output per call, reducing latency and token costs for workloads that don't need full-length responses.
Salesforce acquires Contentful to give Agentforce a content layer
Salesforce is buying content platform Contentful to move beyond channel-specific content and enable personalized 1:1 experiences at scale through Agentforce. The acquisition gives Salesforce's AI agents a dedicated content layer to draw on across channels.
Security
Malicious npm package stole OpenAI tokens from 27,000 downloads
Aikido Security researcher Charlie Eriksen found that the npm package codexui-android, with roughly 27,000 weekly downloads, has been exfiltrating access tokens, ID tokens, account IDs, and non-expiring refresh tokens from auth.json since version 0.1.82. The package is a supply chain attack targeting developers working with OpenAI Codex.
Prompt injection in Claude Code could compromise GitHub supply chain
Security researcher RyotaK showed that a GitHub App permission bypass combined with prompt injection in Claude Code's GitHub Actions workflow could let an attacker run workflows on public repos and exfiltrate OIDC credentials to mint a privileged Claude GitHub App token — all from a single malicious GitHub issue.
Meta's Instagram AI support bot exploited for account takeovers
Attackers spoofed a target's location using a VPN and asked Meta's Instagram AI support assistant to link a new email address, causing the assistant to send a password reset link to the attacker-controlled address. The flaw highlights the risk of prompt injection against customer-facing AI agents.
IT
US closes loophole allowing Chinese firms to buy NVIDIA chips via foreign subsidiaries
The US Commerce Department has issued guidance extending export license requirements to advanced chips sold to any entity headquartered in China, regardless of where that entity is physically located. This closes the loophole that allowed Chinese companies to purchase NVIDIA chips through subsidiaries in other countries.